Session: 01-04: AI for Energy Sustainability IV

Paper Number: 131031

131031 - Sensor Anomalies Characterization and Detection via Machine Learning Methods for Nuclear Power Plants 

Abstract: 

The increasing complexity of next-generation nuclear reactors presents significant challenges in regard to effective system parameter monitoring and qualified data management, as well as the avoidance of anomalous data and noise. This makes it essential to differentiate anomalous sensor data from typical noise and expected data patterns. Current models, such as ones employing digital twins, show promise, but there is a need for fundamental analyses conducted based on the use of simplified models. The present study proposes a foundational approach in which simulated data from PCTRAN-generated datasets are utilized to emulate reference pressurized-water reactor (PWR) system steady-state conditions, and in which typical sensor performance and unusual sensor anomalies are introduced to these datasets. The developed method focuses on data partitioning and linear regression for preprocessing, and utilizes a K-means clustering algorithm for anomaly detection. The initial results exhibit high accuracy, with most instances showing over 95% precision in identifying anomalies, except for those that reflect exponential drift characteristics. A parametric and sensitivity analysis conducted using the Risk Analysis Virtual Environment (RAVEN) tools, developed at Idaho National Laboratory. Parametric study using Monte Carlo Sampling is performed on the anomaly detection algorithm. The input values of R2 tolerance, slope tolerance and window size—defines how many data points each data sample contains—are perturbed randomly within ranges [0.01,0.1], [0.05, 0.5], and [5,20], respectively. The Monte Carlo sampling strategy is used to generate a complete set of options for the input parameters and the “accuracy” and “time” values have been determined for each option. Pareto optimal frontier, an envelope of options that dominates (in terms of both accuracy and time) the set of remaining options has been determined through RAVEN ParetoFrontier postprocessor. The primary objective of this analysis is to identify the optimal option for the input parameters so that the “accuracy” is maximized in a manner that the “execution time” is minimized. Once the Pareto frontier is identified, one can also impose some constraints on the “accuracy” and “time” so that the selected points satisfy users provided requirements. To consider the impact of uncertainties in the input parameters on the performance of anomaly detection algorithm, we have performed sensitivity and uncertainty analysis on one of the optimal options. Result generated insights into the sensitivity of various input variables, highlighting window size as being the most critical factor affecting accuracy and computational time. Although the models require additional refinement for practical application in nuclear systems, the positive outcomes underline the potential value of these techniques.

Presenting Author: Palash Bhowmik Idaho National Laboratory

Presenting Author Biography: Dr. Palash Kumar Bhowmik is a staff scientist (former post-doctoral research associate) in the Irradiation Experiment Thermal-Hydraulics Analysis department (C140) at Idaho National Laboratory (INL) with a research focus on reactor system thermal-hydraulics design and analysis. Dr. Bhowmik’s academic background includes a masters and a Ph.D. in nuclear engineering, and an MBA. Dr. Bhowmik’s industry experience also include providing oversight and approval for procurement specifications, design analyses, design basis calculations, system modifications, commissioning, and closeout activates. He received professional development training from organizations such as the International Atomic Energy Agency (IAEA), Argonne National Laboratory (ANL), and Oak Ridge National Laboratory (ORNL). At present, Dr. Bhowmik is mostly involved with the development of test facilities for the advanced small modular reactor (SMR) system under the United States (U.S.) Department of Energy (DOE) Advanced Reactor Demonstration Project (ARDP). He holds affiliations with the Institute of Electrical and Electronics Engineers (IEEE), the American Society of Mechanical Engineers (ASME), and the American Society of Nuclear Engineers (ANS).

Authors:

Liam Pohlmann University of New Mexico
Palash Bhowmik Idaho National Laboratory
Congjian Wang Idaho National Laboratory
Piyush Sabharwall Idaho National Laboratory